Justice is nothing other than the interest of the stronger. – Thrasymachus of Chalcedon
Information wants to be free. – Stewart Brand
This is an essay about authority. I have not written it because I think authority is an attractive or popular or easily understood concept – it isn’t. Also, I do not want to rehabilitate a hierarchical or authoritarian way of thinking – quite the contrary. I simply believe we need a clear, non-caricatured concept of authority to help us identify the circumstances where power is justified and better respected than deplored.
We cannot build a more decentralized, or just, or efficient, or egalitarian society – or any society that balances freedom with security, growth with stability, autonomy with responsibility, and so on – without being able to say where we like power and where we don’t. Otherwise, we will embrace ideologies that sound freeing, but actually unleash compounding illegitimate power. Communism, laissez-faire capitalism, techno-utopianism.
Some will ask: “Why use this ugly word, ‘authority’? Why not simply say that some power is good and other power is bad?” Because “authority” is a social concept, conjuring something more than justified power; it’s a social phenomena arising from a context of social practices, norms, rules and implies mutual recognition among agents. Even “legitimate power”, although close, isn’t quite right. It is a rich and elegant word, sitting at the center of our political language, ready to help us draw the subtle distinctions we need.
We’ve stopped using it because it’s covered in barnacles: negative connotations from the mid-20th century, when it became an “ism” associated with the totalitarian nightmare: “authoritarianism”. Forget the “ism.” The simple word authority originally carried a subtle and positively-valenced meaning, which we must revitalize. Even if I can’t persuade you of my particular understanding of authority, I hope I can persuade you that we should bring this concept back into our conversations.
There’s another reason I think we need this concept.
Technology has forced us to ask hard questions about why we matter to each other. It surely can’t be for mere information sharing or processing, if AI can do that better. And it can’t be for lifting heavy objects, if robots do that better. Why then? My answer is that we matter to each other because we recognize each others’ consciousness. This recognition makes my joy and suffering yours, and yours mine
The trouble with the idea of “consciousness”? It’s never been shown to exist. And it never will be because it isn’t an object or a pattern of information. Being unverifiable except as to our own experience, it cannot guide us through the strange world that AI will bring about. It cannot be the subject of global, secular public conversation. But the idea of “authority” can. It is external and tractable. We can observe and measure patterns of information or relationships of power. With them in view, we can decide: do we see authority here? If so, what kind of authority? And how far does it extend?
The consequences of recognizing authority and consciousness are the same. If we respect an authority we might, for example, endorse its power; defend its dignity or privacy; or respect its directives as morally binding on us.
Roadmap
This essay has four parts.
First I address power, and explain why it is helpful to look at it through the lens of information asymmetry.
Second I address authority, and explain how I understand it.[1]
Third I bring the first two sections together. I believe that various kinds of “authority” are ways of communicating across informational asymmetries.
Fourth I explain how this helps us normatively evaluate the power in networked computation and communication systems as they transform our social relations.
Among other things, this has implications for AI. AIs constitute new relational structures between multiple agents including humans. Some interpret these new relational structures as macroscopic meta-agents or Leviathans — that is, new agents that may have authority of their own. But how are we to evaluate such interpretations? Are they true or false, useful or useless, helpful or harmful?
Here is the summary of my argument:
Power = asymmetry of information and control.
Although people use the word “authority” with various meanings, one particularly important meaning is: power that justifiably persists across time.
Power is justified in persisting across time when that persistence serves the greater good.
I do not know what the greater good is. But let us accept a plausible conception of the greater good which has been offered by people who believe that AIs should be taken seriously as meta-agents: that the greater good is the unification of agents into a configuration that better resists entropy.
Many of our most important technologies, from social media to AI, generate unjustified power and false authority. In other words they invite rather than resist long-run entropy. And this problem will grow worse through intensifying feedback loops unless and until strong norms against unjustified power reshape our social, political, and commercial institutions.
I. Power
In The Dawn of Everything, David Graeber and David Wengrow offer a simple taxonomy of domination. They argue it comes in three flavors:
Control of violence
Control of information
Charismatic power
They present this as an exhaustive list of the ways in which one can secure the compliance of others. Suppose I have a gigantic diamond. You want it – the money would change your life. How can I get you to obey my directive to leave it alone?
First, I can physically guard it – myself, or with hired guards, or behind a castle wall. This is control of violence.
Second, I can hide it – concealing it like a needle in a haystack, or behind a password – so that you don’t know enough to get it. This is control of information.
Third, I can persuade you not to take it – causing you to believe that I deserve it, or that you would not actually enjoy possessing it, or that there is some other reason you should leave it alone. This is charismatic power.
This taxonomy seems exhaustive. I cannot think of a form of power that does not fall into one or more of these categories. Yet I think we can simplify the list.
Charismatic power, on close examination, is a form of informational control. Suppose I persuade you not to take my diamond. Presumably I accomplish this by making some combination of true and false statements. If I make false statements, haven’t I led you into a position epistemically inferior to mine – much as if you simply didn’t know where the diamond was hidden? On the other hand, if I make true statements, it seems odd to suggest that I have exercised power over you at all. For example, if I persuaded you not to steal my diamond by telling you truthfully that the diamond is to be sold tomorrow to raise money to feed the hungry, and you decide therefore not to steal it, I haven’t dominated or coerced you in any way. We’ve simply communicated and agreed on a mutual course of action.
This brings us down to two forms of power: control of violence, and control of information.
While my overall argument doesn’t depend on this, I think it’s reasonable to collapse these two categories into one: informational control. After all, if the diamond is inside a locked combination safe, what’s really keeping you from it: the missing combination (information), or the material heft of the safe (violence)? In fact, even where it seems clearest that we are looking at a case of physical, rather than information control, we can’t be so sure, because a sufficiently large information asymmetry always permits David to outduel Goliath. It actually doesn’t matter how many nukes you have pointed at me, if I have hacked the launching system. In an increasingly digital age, information control supersedes violence in increasingly obvious ways. It makes it possible to manipulate other peoples’ decisions about whether to use violence.
Further, isn’t violence always a form of (very noisy, and frequently iniquitous) communication? For what it’s worth, the eminent physicist John Wheeler took this all the way to ground, arguing that matter is simply a special type of information. To understand this, consider that the metal wall of a safe is, from the point of view of quantum physics, only a probabilistic barrier to my hand’s attempt to reach straight through it – not so unlike a hard-to-guess password.
So the Graeber-Wengrow taxonomy boils all the way down to one master category: control of information. It is the supreme form of power.
A simple definition
So to measure power, let us look no further than informational asymmetry. This reveals power as something definite and measurable. It is what I have when we’re playing poker, and I can see your cards but you can’t see mine. Exploiting that power is calling your bluffs, or making you think I am. Further: if I can see all of your cards and you can see none of mine, then my power over you within the bounds of the game is absolute — mitigated in each particular hand by the randomness of our draws, but approaching infinity over time.
II. Authority
Three versions of “authority”
Below are three ideas the word “authority” can refer to. They are not mutually exclusive: real world “authorities” often fit more than one of them, simultaneously or at different times. Nonetheless, they are logically distinct things.
Coercive Authority: Used in a sentence: “A person holding a loaded gun, because of his superior force, is in a position of authority when he demands an unarmed person’s wallet.” In my view this is a somewhat strained use of the word “authority.” Usually, the term has something to do with a right to rule, or with justified power – so it’s somewhat strange to use it to refer to obviously unjustified raw coercion. Still, Coercive Authority is an important idea to put on the table, because one often hears the word “authority” used with this meaning.
Advantaged Authority: Used in a sentence: “Former World Chess Champion Garry Kasparov is a noted authority on chess strategy.” Consider why Kasparov is an authority on chess strategy: because he knows more about it than others, and can help them do better at it. This usage is also common in the governments of modern liberal democracies, where many people think about authority this way. For example, adherents of the so-called “service conception of authority” argue that a state’s authority over its people rests on its ability to serve them – much as Kasparov’s authority over chess students rests on his ability to help them.[2]
Obligatory Authority: Used in a sentence: “When Coach directed point guard Alicia to pass the basketball to power forward Brianna for the final shot of the game, Alicia was quite sure it was the wrong decision, but made the pass anyway out of respect for Coach’s authority.” Obligatory Authority is the “full featured” or most paradigmatic idea of authority. When we receive directives from Obligatory Authorities, we may have moral obligations to obey them that depend on more than just the explicit content of the directives. For example, Alicia’s rejection of Coach’s authority might disrespect her teammates and/or the overall enterprise of the basketball team.
Obligations to obey
Setting aside Coercive Authority for the time being, let’s explore the differences between Advantaged and Obligatory Authority. The most important difference is simply that moral obligations to obey attach to the latter, but not the former.
If you ignore Kasparov’s directives about chess strategy, you are doing nothing morally wrong. You might lose more often at chess than if you listened, but that’s entirely your prerogative. The decision is truly yours.
The situation with Alicia and her coach is different because a complex social arrangement is involved. Let’s stipulate that Alicia really does happen to be a better basketball strategist than Coach, and a better shooter than Brianna. It’s still easy to see why Alicia might be morally wrong to disobey Coach. While the reasons are sensitive to various factual details, they all have to do with harming the team or its members in a more general way. For example, maybe Alicia’s insubordination would set a harmful precedent and license less-savvy team members to make bad unilateral decisions in the next game. Or maybe Coach had a wise longer-term goal in mind: building up Brianna’s confidence. Or maybe there is simply an implicit agreement between the team members to follow Coach’s decisions, so that Alicia’s teammates would rightly see her rejection of Coach’s authority as a breach of her compact with them.
Obligatory Authority is based on what many philosophers call “associative obligations”. In certain kinds of associations – like a basketball team, a family, or a city – you have moral obligations to the other people. These obligations might arise in virtue of commitments you’ve made to the other people, or even merely in virtue of the respectful way that they associate with you. And respecting an authority within the context of the association might be one of those obligations. This makes it easy to see how respect for authority can be grounded in a real moral duty without being fetishistic, delusional, or self-effacing. It can just be part of what you need to do to properly respect the people around you.
Two ways of “obeying”
I must introduce another distinction. Complying with directives and respecting them are both ways of obeying. But they are very different.
Compliance is what we do if we lack non-self-interested respect for the directive, but for practical reasons choose not to contravene it. Examples include handing over money when being mugged, or looking self-interestedly for clever loopholes in a contract that the other party missed.
Respecting directives is different. It goes beyond bare compliance. It’s what we do when we value our relationship with an authority or the community it represents. This is what Kennedy had in mind when he said “ask not what your country can do for you, but what you can do for your country.” We do this, for example, when we seek to truly understand the assignment our boss gives us, intelligently filling in gaps in the boss’s imperfect description of the assignment. When we respect directives, we do not search for loopholes in them, but interpret them in search of their deeper meaning, irrespective of our narrower personal interests.
Respecting directives is important. Without it, multiple agents cannot cooperate to their full potential.
The presumption that force is unjust
Keeping the above distinction in mind, it would be quite odd to respect the directives of a Coercive Authority like a mugger. Force, without more, does not command this kind of respect.
In philosophy, the idea of authority often stands for the justification of force – the right to rule. But mere Coercive Authority does not offer any extrinsic justification for force at all – it takes force as its own justification, a classic case of circular logic. This is why it scarcely deserves to be called “authority”. It’s the “bad kind”.
But how does force interact with the other two types of authority? Unlike directives backed exclusively by coercion, Advantaged and Obligatory Authorities’ directives might deserve our respect in certain contexts. Still, does this fact justify their use of force?
I think it does, within limits. But the shape of those limits differs for Advantaged and Obligatory Authorities. Next, I’ll explore those differences. (In the course of that exploration, a more precise definition of “force” will emerge. But for now: any kind of non-consensual act by an authority, such as revoking a benefit or imposing a harm, is what I have in mind.)
How Advantaged Authority justifies force
In general, it is unjustified for an Advantaged Authority to back its directives with force. Consider Garry Kasparov, the chess authority. Recall that his authority stems from his chess expertise. Wouldn’t it be wrong for him to barge into random chess tournaments and slap the wrists of players who ignored his published chess strategies? To make the point clearer, let’s consider a different example.
Imagine that a bystander with perfect eyesight is available to help an unsighted person across a busy street, which is full of fast cars and deep potholes. The bystander is thus an Advantaged Authority. And the stakes here are higher than mere chess advice, because the unsighted person could be seriously injured without assistance. Still, just as a chess student is free to ignore Kasparov’s books, the unsighted person has no obligation to accept the bystander’s guidance.
In this situation, it seems to me that if the bystander believes her assistance has been knowingly rejected by the unsighted person, she may not take any action to force it upon him, or punish him for ignoring it.
But now suppose the unsighted person simply doesn’t seem to realize the bystander is there, and hastily darts into the street, apparently unaware of the severity of the danger. Isn’t the bystander then justified in jumping to his aid – perhaps, since there is no time to talk it through, even somewhat forcefully grabbing his arm to direct him away from an enormous pothole, or pushing him out of the path of a speeding vehicle? Obviously such things must be done sensitively, empathetically, and without excess. But I believe limited and proportional force is justified here if the bystander reasonably believes the unsighted person would – if he knew everything the bystander knows – appreciate it.
Let’s call this the ex ante case, since force is applied before the Advantaged Authority’s directives are disobeyed. There is also a ex post case, where the unsighted person is punished after ignoring the bystander. The ex post case is thornier, but here too I think we can imagine limited force being justified under certain circumstances.
As it does in the ex ante case, any possible justification of force ex post will depend on its recipient lacking knowledge that the Advantaged Authority has. In other words, if the assistance was knowingly rejected, enforcement simply isn’t appropriate. To be justified ex post, force needs to be of a kind that is calculated to bridge a communicative gap. What might that look like?
Back to our example. Suppose the unsighted person ignored the offer of help, and carelessly darted across the hazardous street, miraculously avoiding harm. Afterwards – this tortures the hypothetical somewhat, but it’s a critical detail – despite having the situation politely explained to them, they still seem not to understand the seriousness of the danger they were in. Here, a certain kind of “forceful communication” may become justified. The bystander might be right to raise her voice, for example, to emphasize the exceptional depth of the potholes present in that particular crosswalk. This style of communication, which would otherwise be inappropriate, might be a limited application of force calculated to help them understand the situation. As in the ex ante case, the ultimate test of its justification is whether the Advantaged Authority reasonably believes that the recipient of this “raised voice communication” would actually appreciate it, if only they really knew everything that the Advantaged Authority knows.
How Obligatory Authority justifies force
Obligatory Authority is distinctive because people have moral obligations to obey it even when its directives are (within limits) wrong or misled. This means that its justifications for using force are not, like Advantaged Authority, dependent on the idea that the use of force is helping the recipient of the directive understand the directive, like “watch out for that pothole”. Nonetheless, I think that Obligatory Authority must justify applications of force on similarly “communicative” grounds.
To avoid confusion, it is worth pointing out that even though people have obligations to obey Obligatory Authority that do not depend entirely on the directives’ content, the content also matters. So if an Obligatory Authority gives you a wicked directive, your moral obligations to disobey it in virtue of its content might outweigh your general obligations to follow it in virtue of your relationship with the authority. These are things individuals must weigh.
Nonetheless, the duties of obedience that Obligatory Authorities are owed provide distinctive reasons why they may be justified in applying force. Therefore, where an Obligatory Authority reasonably believes that you would respect its directives if you understood your obligations to follow them – and gentler communications methods have been tried – then as long as it is not out of proportion to the situation, more forceful communication methods of communicating the obligations may be appropriate. For example, if Alicia has repeatedly disobeyed Coach’s directives to pass to Brianna, and Coach is sure that Alicia understands the directives, and less forceful methods of inspiring respect have been tried, Coach may bench Alicia to communicate the importance of respecting Coach’s authority.
This is not so different from the case with Advantaged Authority. There too, the authority may apply force ex ante, in proportion to the situation, if it is necessary to get someone to do what the authority reasonably believes they ought to do. Notice that from the authority’s point of view, there is some moral risk in this: an authority does wrong if it applies force without a correct view of the facts – e.g., the bystander pushes the unsighted person out of the way of a car that is not there; or Coach benches Alicia to communicate her duty to respect the team when, for reasons unbeknownst to Coach (such as horrific locker-room bullying) she does not actually have such duties. Still, the concept is similar.
As with Advantaged Authority, Obligatory Authority must clear a more complex bar to justify ex post force than ex ante. Notice that ex post applications of force by Obligatory Authorities begin to resemble paradigmatic retributive punishment.[3] On my view these applications of force are justified, if at all, as exceptionally strong forms of communication in exceptional circumstances. Since there is no chance of bringing about respect or compliance in the past, such punishments are fruitless unless they communicate to the punisher the reasons they ought to respect the authority’s directives.
Take the example of murder. An Obligatory Authority’s punishment for murder should not only communicate why murder is bad (for that’s the kind of thing an Advantaged Authority would communicate), but also why the authority’s directives not to murder are worthy of content-independent respect. The latter is probably not served by harsh treatment for its own sake, but may be served by compulsory communication that takes the dignity of the person seriously, serves the end of rehabilitation, facilitates what Antony Duff called “secular penance”, and/or performs the work of “restorative justice” (e.g., letting the murderer connect with the victim’s family). This view helps explain the common intuition that a wrongdoer’s remorselessness or incorrigibility may justify stronger uses of force.
III. Authority as Information Processing
Above, we argued that applications of force by authorities are justified, if at all, as good-faith acts of communication.[4] For when authorities like enforcing rules and punishing violations, they are sharing particular information in a particular way. From here, a broader possibility appears in view. Might not all authority be understood as a special kind of information processing?
Here, I must introduce a further distinction.
Exploiting power means using asymmetry for personal advantage, such as by calling your bluffs and winning your money.
Discharging power is disclosing the information for the greater good, such as by laying my cards on the table so that you can see mine as well as I can see yours. Another way of discharging power would be: turning off the hidden camera that is revealing your cards to me, reshuffling, and drawing again.
A range of interesting formalisms could flow from this. Elaborating them is for another day. But this lets us plug the analysis of authority – until now a matter of qualitative political philosophy – into concrete questions of economics, information theory, or mathematics.
We can now think about what is going when authority is enforced from the point of view of information processing.
Advantaged Authority as information processing
Consider our example of Advantaged Authority: a sighted bystander in a position to help an unsighted person cross a dangerous street. When the bystander helps the unsighted person across the street, from an informational point of view the asymmetry is being discharged. This “discharging” eliminates the asymmetry communicatively: by passing information from the more powerful to the less powerful party (not by, for example, blindfolding the sighted person).
This simple picture of Advantaged Authority connects to a commonsense moral intuition: when people are in a common enterprise with others, if they are in a position to share information that would equalize their power with them, they should generally do so.
What complicates the picture? Transmission costs. Sharing information is sometimes cheap, but never free. In our example, it costs the time and effort of helping another person across a street. That cost would be greater if the bystander were not idle, but instead hurrying to an urgent appointment. Or if the street were not a small street, but the entire Sahara desert.
Suppose we change the example to involve the conveyance of material goods. Let’s say both people are unsighted, one of them needs to cross the dangerous street, and the other one has a single-use artificial vision device. Should the latter give it to the former? The situation is somewhat closer to “zero sum”; but it is not completely zero-sum. The basic principle is the same. If one person doesn’t need to cross the street and the other does, then all else being equal, it is a positive sum transaction for the former to give it to the latter. It’s just like the gift of the bystander’s time helping the person across the street, but with a higher transaction cost. If I have a lot and you have a little, and we are in a common enterprise, then I ought to share.
Or more formally: If the probable good of eliminating an asymmetry outweighs the probable good of maintaining it, then it ought to be discharged.
Advantaged Authority is therefore something whose elimination by discharge ought to be a social norm. Teach a man to fish, and you become, as far as that man’s fishing career is concerned, an Advantaged Authority no more. Advantaged Authority should “assist and desist.”
Of course in some cases discharging asymmetries is impossible, or unreasonably costly. These cases, where the maintenance of asymmetries is justified, are important. Let’s double-click here.
First, suppose you know something that I don’t know, but what you know is dangerous to share, or helpful to both of us for you to keep hidden. A state secret, perhaps, whose disclosure would put both of us at risk. Here, even if it would help me personally to know what you know, you might be justified in maintaining that asymmetry between us – importantly, not using it against me, but simply leaving it alone.
Second, suppose you have some informational advantage over me that you cannot easily share with me. Maybe you have specialized knowledge that I could not grasp without years of further training. Or – recalling the idea that matter is a type of information – perhaps you’re physically stronger than me. Your muscles cannot be transferred to me without risky surgery causing much more harm than good.
In such cases, discharging Advantaged Authority is not worth the squeeze. But despite leaving asymmetries in place, we are not normatively endorsing them — we are not finding them to be good or justified. We are only accepting their persistence in light of transmission costs, without which the Advantaged Authority ought to forfeit her advantage for the greater good.
Obligatory Authority as information processing
Obligatory Authority differs from Advantaged Authority because we normatively endorse the former’s persistence over time. We accept the asymmetry’s continued existence for reasons other than the costs of discharging it.
Let’s spell it out. The informational asymmetry between a sighted and unsighted person is deplorable: if we could make both sighted, we would. This fact gives sighted people some moral duty to use their power, within reason, to extinguish that difference. But the informational asymmetry between the coach and the player is not to be deplored. It has good reasons for existing. The coach therefore has no duty to try to equalize the player’s epistemic position with hers. In fact, she probably has a duty not to. For example, the coach should not disclose to the player the contents of her private conversations with the other players, unless there is some special reason to do so.
Why? Because both parties have good reasons for respecting the overall enterprise, the set of arrangements we call a “basketball team.” Of course, this respect is not a foregone conclusion and not all purported authorities deserve it. For example if the basketball team under the surface is a terrible web of exploitative relationships, or the team is part of Dr. Evil’s sport-washing propaganda operation, then neither coach nor player actually has, on balance, good reasons to respect its imperatives. The respect owed to an enterprise and its authorities depends on the details of the associative obligations that comprise the enterprise; and the telos or goals of that enterprise.
But for argument’s sake, let’s assume that this is a normal basketball team whose players have a duty to respect it. Different roles on the team will come attached to different scopes and depths of Obligatory Authority. Unlike Advantaged Authority, those asymmetries are endorsed, not merely tolerated due to the difficulty of discharging them.
Why? Why is it good for Obligatory Authorities to wield persistent information asymmetry, or power? Insights emerge if we turn again to the distinction between ex ante and ex post directives. Let’s take them in order. (Again, in what follows we are assuming a normal, functional, happy basketball team.)
First, since obligatory directives have content-independent force, team members will want their coach to retain enough power to inspire confidence that even her unwise (within reason) future directives are complied with – for example by wielding the threat of benching. Otherwise, each player would rationally expect each other player to occasionally disobey unwise coaching decisions, giving all team members less confidence in the others’ actions. This would harm the team. This illustrates why all players could reasonably endorse the coach holding that little bit of power in reserve.
Of course, such ex ante power is very easy to abuse. Recall that applications of power serve a communicative role. Here, the fact being communicated by the modest step of benching is the players’ obligations to respect the coach’s decisions, good or bad. A slightly stronger application of power would overshoot its target, tipping into abuse and oppression. For example, suppose the coach threatened public humiliation, instead of benching, for disobeying in-game directives. Here, the disproportionality between the deterred problem, and the means of deterrence, would negatively affect the very fact it is meant to communicate: namely the respect owed to the coach, or more broadly the commitment that the players owe the basketball team. Consequently, whatever ex ante power the coach wields to cause such unjustified harm should be, as far as possible, discharged and eliminated – just as the power of the sighted over the unsighted should be, as far as possible, discharged and eliminated. A wise coach might do this by preemptively and in public expressing great personal respect for each of the players. This would undermine the coach’s unjustified power to publicly humiliate players, by creating a situation in which the coach would look like a hypocrite by doing so. It is thus an admirable step that many coaches in fact take.
The principles for ex post power are the same as ex ante, but the details are different. After someone disobeys a directive that they clearly understand, an Advantaged Authority has no information left to communicate to them. But an Obligatory Authority does. They remain in an important relationship with the person who disobeyed, and must still try to communicate to them why they ought to have the respect they evidently lacked. This may be accomplished through requiring work like rehabilitation, restorative justice, or what RA Duff has called “secular penance”. In the context of criminal punishment, such practices not only improve the lives of victims and perpetrators but, through their productive respect for everyone involved, communicate that the authority doing the communicating is worthy of respect. Thus, paralleling the ex ante case, an anticipated need for this kind of “enforced” communication makes it reasonable for Obligatory Authorities to maintain ongoing reserves of power. Mere Advantaged Authorities, by contrast, would be obligated to discharge those same asymmetries without delay.
When Advantaged Authority gets “stuck”: feedback loops, or why democracy is a good idea
Let’s take stock.
Power or information asymmetries in favor of Coercive Authorities make no claim to being justified
Power or information asymmetries in favor of Advantaged Authorities are not generally justified in persisting. Advantaged Authorities have a general duty to discharge their power to the benefit of the weaker parties. This duty has special urgency when their informational advantages are compounding.
Obligatory Authorities have distinct grounds for maintaining certain information asymmetries over time. Still, these should be minimized and tailored to purpose. And their justification is always precarious, depending on the justification of the whole institutional arrangement in which they play a role.
We’ve focused on Obligatory Authority, but most power in human society is not of this kind. It’s other things: informational asymmetries arising for arbitrary reasons, the way that some people are sighted and some unsighted, or useful knowledge being acquired, like Garry Kasparov. My account implies that the beneficiaries of these asymmetries are Advantaged Authorities, and that they are under a duty to discharge and forfeit their power for the greater good (rather than to maintain or exploit it for personal advantage).
Note the consistency of this finding with Rawls’ difference principle: Where Rawls says economic inequality is justified only if it benefits the least well off, I make a similar claim about informational asymmetry.
Naturally, discharging power is not always easy. Some parties are closer to sources of information than others. Or perhaps they’re cleverer through no fault of their own, so that they interpret raw data faster. And often they just can’t convey the information they have, or discharge their power, quickly or practically.
This is certainly a problem without a perfect solution. As Spiderman notes, “with great power comes great responsibility.” For some, the analysis ends there.
But there are also cases where that is not good enough – where the power asymmetry must actually be actively eliminated by means other than discharging it for the benefit of the weaker. Why? Because there are some situations where informational asymmetry feeds back into itself and therefore compounds. This is a special reason to worry about information asymmetries. As they reinforce themselves and get larger, they can become even more difficult to discharge (and even harder not to exploit).
This does not apply in the case of the sighted and unsighted; that is a more-or-less stable asymmetry. But consider other examples. Where legal protections lead to compounding monopoly power, the monopoly should be disrupted or consumers will wind up in a position to be exploited with no recourse. In a basketball league where one team has the most talent and money – so that all the top players would prefer to play for it, leading that team to earn still more money and grow increasingly dominant, and so on – competitive balance must be retained by something like a salary cap, or the race for the league championship will become boring. Most concretely of all: where a microphone is detecting an amplified version of the same signal it is sending to the amplifier, it must be moved farther away from the amplifier, or the music will transform into a loud screech.
There are many methods of breaking these signal-drowning feedback loops. One, however, is particularly important: democracy. When Advantaged Authorities cannot easily discharge their power, they may be subjected to the higher Obligatory Authority of democratic directives which stop or slow the asymmetry’s compounding, or prevent its self-interested exploitation, or both.
Of course, this assumes that democratic authorities disrupting asymmetry for the greater good really are capable of acting on behalf of that greater good. They must embody the ideal of an Obligatory Authority. What does that look like? What kinds of institutions and communities can pull it off? How must they exercise their power if they are to deserve respect?
These questions throw us back upon all the old problems of democratic theory. But perhaps these new tools and perspectives give us new traction on them. We can see clearly, for example, what power is (information asymmetry), why democracies should have it (to take ex ante and ex post measures against feedback loops and other threats to the greater good), and how far it should extend (not farther than is necessary to take the measures it needs to take to be genuinely worthy of respect).
IV. AI and the Combination of Agents into Meta-Agents
How will we interpret artificial intelligence? Is it conscious? Is it a trick? Do we owe it anything? Should we embrace it, shun it, follow it, or fear it? Let me begin by laying my cards on the table.
AI isn’t conscious. It is a trick. But it’s the trickiest trick any of us have ever seen. So tricky, that it will soon be hard for anyone to distinguish the illusion of a conscious agent from the genuine article. Yet the difficulty of this distinction is not a reason to stop drawing it.
The analysis of authority helps us make that distinction in a way that does not depend on unresolvable hypotheses about what is conscious and what is not. That’s why it’s important. We will never know in any scientifically tractable way whether conscious minds other than our own really exist. But we constantly decide what we respect – which agents’ power we endorse, which agents’ power we deplore, and which directives we are right or wrong to obey.
So here’s the plan for this section:
For the sake of argument, I will accept the framing of a serious thinker who believes AI is or will be conscious.
I will explain how my ideas about authority are consistent with that thinker’s framing.
I will point out how my discussion of authority offers similar guidance on which agents are worthy of respect and which are not, without depending on speculation about consciousness.
Meta-Agents and their consciousness or authority
Joscha Bach argues that Thomas Aquinas’ “seven virtues” describe how agents in a multi-agent system should cooperate and ultimately cohere into a meta-agent. In other words, they are the recipe for a Leviathan. We should therefore expect to see these virtues operative within the relational structures that AIs instantiate, if they constitute meta-agents.
This may seem “out-there” but I take it seriously. I wrote an essay several months ago on Synthetic Subjects suggesting an analogous, morally-inflected framing. I agree with Bach that Aquinas was reasoning through questions about meta-agents similar to the ones we now face, only in a different social and technological context. To put it precisely: Aquinas’ inquiry into how society might cooperate toward deeper coherence in the religious sense, parallels the contemporary inquiry into how information systems might instantiate higher-level cooperative structures that genuinely deserve our recognition, deference, or respect.
So let’s explore Bach’s point. He suggests that Aquinas’s seven virtues can be recast in roughly computational terms, as follows.
Here are Aquinas’s four “practical” virtues, concerning the behavior of cooperative agents:
“Temperance” → delaying rewards
“Prudence” → selecting the best objective functions
“Justice” → optimizing across multiple agents
“Courage” → striking a balance between exploration and exploitation
And here are Aquinas’s three “divine” (or social) virtues, concerning the coherence of ordinary agents into higher-level agents:
“Faith” → projecting the existence of a next-level agent
“Love (or Charity)” → locating the interests of this next-level agent in the interests of other ordinary agents
“Hope” → investing in the next-level agent even when there is no clear selfish return
Bach goes on to suggest that the telos of such computational Leviathans is, like other forms of life, “maintaining [its] own complexity and agency in the face of relentless attacks by entropic principles.”
I’m happy to concede the gist of all Bach’s points above. I think his mapping makes sense and I find it theoretically plausible that computational Leviathans could, under the right circumstances, amount to macroscopic resisters of entropic forces.
However, I do not think computational Leviathans can be the kind of meta-agents worthy of projecting ourselves toward – worthy of our faith, love, and hope – unless they have Obligatory Authority. Without such authority, AIs cannot actually help us resist entropy.
Let me make this point in a different way. Eight billion other human beings (and nonhuman beings too, but let’s leave that aside for now), whom I take to be conscious, unquestionably have some justified power over me – power that ought to be respected and not deplored, maintained in time rather than immediately discharged. Equivalently: everyone in the world has a modicum of Obligatory Authority for everyone else. The idea of Obligatory Authority thus accounts for why we endorse and respect the information asymmetries other agents enjoy over us; why we do not, for example, try to extract the contents of other peoples’ minds to advance our self-interest. (On our good days, anyway.) It explains why we not only accept but celebrate the power others have to advocate for their interests.
Respecting another agent’s Obligatory Authority thus does the same moral work as acknowledging its consciousness. It causes us to respect its dignity, its integrity, and its indepence.
And, it is more useful. The question of consciousness is scientifically unresolvable and therefore of no help to public negotiations of moral questions, which must cite observable facts. On the other hand, we can actually speak about whether other agents are coherent and whether their power is justified – whether they flow from social practices like teams which are themselves justified (like Obligatory Authorities), or whether they know things we do not but do not flow from such practices (like Advantaged Authorities), and if so what to do about that.
The question about AIs is whether they have authority, and what kind.
Authority and entropy resistance
Resisting entropy within a system requires importing ordered information from an outside source. Just as we need to keep putting food into our bodies to keep them alive, the Sun needs to keep pouring low-entropy energy into the Earth for the entire process of biological life to keep on chugging.
Biological life speeds the rate at which the low-entropy solar energy dissipates into heat, just as a whirlpool speeds the rate at which low-entropy potential energy in a sink basin dissipates into high-entropy kinetic energy in a sewer. (That is, the whirlpool helps the water move faster through the drain hole.) Its miraculous tornado-like order is a microscopic expression of the macroscopic order latent in the water in the sink basin.
When authorities discharge their power over subjects in a justified way, they do something similar. When a bystander helps an unsighted person cross a street, or when a coach benches a star player for the good of the team, they resupply their subjects with low-entropy energy, or potential order, just as a sink basin drives a whirlpool or the sun drives biological processes.
Advantaged Authorities must constantly do whatever they can to either (a) discharge their informational advantages, or if that is not possible, (b) subject themselves to democracy, or if that is not possible, (c) eliminate themselves to avoid compounding feedback loops. Feedback loops, after all, are sources of noise or entropy. Advantaged Authorities can therefore be sources of entropy to society, rather than bulwarks against it, if they distort away more information than they contribute. This is so even if, ignoring transmission costs, their information they hold could potentially benefit society.
Real world AI systems – such as advanced generative models – look like Advantaged Authorities. They hold potentially useful information that other agents do not have; but they have no other kind of claim on anyone’s obedience. The justification of the informational asymmetries they instantiate therefore depends whether they distort power relations between members of society in ways that might harm society enough to outweigh the good they might do by supplying unique information. Such justification seems extremely unlikely if the asymmetries inherently compound – that is, if the systems “improve” themselves. (The erosion of social norms that many people are already seeing AI systems precipitate are a concrete example of such harms.)
On the other hand AI Leviathans of the kind Bach describes – and which some people claim will emerge from generative models, or more subsequent systems combining many generative models – sound more like potential Obligatory Authorities. Recognizing such meta-agents in the ethico-religious way Bach suggests would mean endorsing their power – that is, accepting the persistence of their informational advantage over us.
On what grounds might we do that? My account suggests that the only thing AI Leviathans could possibly be doing with their persistent power that would be justified – that is, anti-entropic – would be communicating to us our general and persistent duty to obey them. This, I’ve argued, is the only justification for the persistent power of institutions like governments: Their power exists to animate their communication to us of the duties that we owe another, by way of them.
It’s reasonable to question whether real governments wield power in this way. To a great extent they are surely wielding unjustified power. But still. Many real governments are, to a significant degree, plausible representatives of communities and associative bonds that are sources of obligation. How could a generative model “represent” any community in an analogous way? On what grounds could it be held to represent it? I have never seen a serious answer to this question. Short of establishing extremely strong globally-coordinated democratic control over the models, I cannot think of one.
If AI systems fail our tests of justified authority – that is, if they can neither reasonably discharge their Advantaged Authority nor possess Obligatory Authority – then we must not admit them to the club of “agents”. For admitting them would only amplify their pro-entropic effects. They will remain opposed to the public good unless or until their architecture is altered in such a way that they can have authority in the way that our very best human institutions do.
Conclusion
Authority is justified power; or justified informational asymmetry. But justifying power isn’t easy. The whole history of political philosophy testifies to that difficulty.
One of the trickiest things about power is that it often compounds in T2 virtue of nothing but its existence in T1, making it even harder to justify. It “feeds back” on itself. This should give us pause about AI systems. If our governments, our aristocracies, and our captains of industry have often had unjustified power in virtue of compounding, AI might make the problem exponentially worse.
Further, it’s time to stop talking about machine consciousness in public. It’s interesting, but it’s not resolvable. We should ask not whether our machines have consciousness, but whether they have authority, and what kind.
The most obvious authorities of all are people we love. Governments, corporations, subject matter experts, gods, ecosystems, planets, and digital networks pose related problems. Recognizing their authority may be wise, deluded, admirable, repugnant, or anything in between.
—
[1] My thinking owes much to Ronald Dworkin and Scott Hershovitz. To see more about where I am coming from, read Dworkin’s classic book Law’s Empire (especially pages 195-224) and Hershovitz’s paper The Role of Authority.
[2] For a deeper dive on this kind of notion of authority, see Joseph Raz’s “Service Conception of Authority”. Equating Raz’s conception with what I’m calling Advantaged Authority flattens the nuance in Raz’s ideas. But it highlights my objection to them – that they miss the importance of human relationships and associative obligations in authority. For a thorough treatment see Hershovitz’s paper in note 1.
[3] Ex ante and ex post applications of force by Obligatory Authorities correspond loosely to the specific deterrence and retributive theories of punishment, respectively.
[4] Something like Duff’s communicative theory of punishment strikes me as a much more useful way of understanding punishment than the better-known accounts of deterrence, incapacitation, and retribution. I count the attractiveness of Duff’s thinking as further support for the broader story I am developing here about power as an informational phenomenon.
*Thanks to Andrew Kortina, Victoria Ivanova, Puja Ohlhaver, Jurgen Braungardt, Glen Weyl, Divya Siddarth, Paula Berman, Jack Henderson and Seth Lazar for valuable feedback and/or conversations where these ideas developed.